Fudan University
AI Insights
- The paper proposes a new approach called InfoSculpt, which uses Conditional Mutual Information (CMI) to tackle the challenges of Generalized Category Discovery (GCD). [3]
- InfoSculpt is based on the Information Bottleneck principle and aims to learn a compressed representation that captures the most relevant information for classification. [3]
- Generalized Category Discovery (GCD): A task that addresses the practical challenge of learning from datasets with labeled known classes and unlabeled data containing both known and novel categories. [3]
- The theoretical foundation provided by the paper offers a deeper understanding of the Information Bottleneck principle and its application to GCD. [3]
- The paper provides a detailed theoretical foundation for InfoSculpt, including a derivation from the IB principle to the CMI objective and an in-depth discussion of the core assumption and refined target distribution. [2]
Abstract
Generalized Category Discovery (GCD) aims to classify instances from both known and novel categories within a large-scale unlabeled dataset, a critical yet challenging task for real-world, open-world applications. However, existing methods often rely on pseudo-labeling, or two-stage clustering, which lack a principled mechanism to explicitly disentangle essential, category-defining signals from instance-specific noise. In this paper, we address this fundamental limitation by re-framing GCD from an information-theoretic perspective, grounded in the Information Bottleneck (IB) principle. We introduce InfoSculpt, a novel framework that systematically sculpts the representation space by minimizing a dual Conditional Mutual Information (CMI) objective. InfoSculpt uniquely combines a Category-Level CMI on labeled data to learn compact and discriminative representations for known classes, and a complementary Instance-Level CMI on all data to distill invariant features by compressing augmentation-induced noise. These two objectives work synergistically at different scales to produce a disentangled and robust latent space where categorical information is preserved while noisy, instance-specific details are discarded. Extensive experiments on 8 benchmarks demonstrate that InfoSculpt validating the effectiveness of our information-theoretic approach.
Generalized Category Discovery (GCD) aims to classify instances from both known and novel categories within a large-scale unlabeled dataset, a critical yet challenging task for real-world, open-world applications. However, existing methods often rely on pseudo-labeling, or two-stage clustering, which lack a principled mechanism to explicitly disentangle essential, category-defining signals from instance-specific noise. In this paper, we address this fundamental limitation by re-framing GCD from an information-theoretic perspective, grounded in the Information Bottleneck (IB) principle. We introduce InfoSculpt, a novel framework that systematically sculpts the representation space by minimizing a dual Conditional Mutual Information (CMI) objective. InfoSculpt uniquely combines a Category-Level CMI on labeled data to learn compact and discriminative representations for known classes, and a complementary Instance-Level CMI on all data to distill invariant features by compressing augmentation-induced noise. These two objectives work synergistically at different scales to produce a disentangled and robust latent space where categorical information is preserved while noisy, instance-specific details are discarded. Extensive experiments on 8 benchmarks demonstrate that InfoSculpt validating the effectiveness of our information-theoretic approach.
Why we are recommending this paper?
Due to your Interest in Continual Generalized Category Discovery
This paper directly addresses Generalized Category Discovery, a key interest area for the user, focusing on innovative methods for classification within large datasets. The approach aligns with the user's interest in automated product categorization and knowledge graph construction.
Due to your Interest in Continual Generalized Category Discovery
This paper directly addresses Generalized Category Discovery, a key interest area for the user, focusing on innovative methods for classification within large datasets. The approach aligns with the user's interest in automated product categorization and knowledge graph construction.
Universit de technologie de Compigne
AI Insights
- Ontological knowledge base: a structured collection of concepts, relationships, and instances that represent a particular domain or subject area. [3]
- The model refinement process is inherently iterative, involving multiple cycles of analysis, update, and validation. [2]
- The proposed methodology for developing ontological knowledge bases leverages the capabilities of Large Language Models (LLMs) to automate various tasks, including identifying core concepts, relationships, and instances, providing accurate definitions, and testing queries. [1]
Abstract
Ontological Knowledge Bases (OKBs) play a vital role in structuring domain-specific knowledge and serve as a foundation for effective knowledge management systems. However, their traditional manual development poses significant challenges related to scalability, consistency, and adaptability. Recent advancements in Generative AI, particularly Large Language Models (LLMs), offer promising solutions for automating and enhancing OKB development. This paper introduces a structured, iterative methodology leveraging LLMs to optimize knowledge acquisition, automate ontology artifact generation, and enable continuous refinement cycles. We demonstrate this approach through a detailed case study focused on developing a user context profile ontology within the vehicle sales domain. Key contributions include significantly accelerated ontology construction processes, improved ontological consistency, effective bias mitigation, and enhanced transparency in the ontology engineering process. Our findings highlight the transformative potential of integrating LLMs into ontology development, notably improving scalability, integration capabilities, and overall efficiency in knowledge management systems.
Ontological Knowledge Bases (OKBs) play a vital role in structuring domain-specific knowledge and serve as a foundation for effective knowledge management systems. However, their traditional manual development poses significant challenges related to scalability, consistency, and adaptability. Recent advancements in Generative AI, particularly Large Language Models (LLMs), offer promising solutions for automating and enhancing OKB development. This paper introduces a structured, iterative methodology leveraging LLMs to optimize knowledge acquisition, automate ontology artifact generation, and enable continuous refinement cycles. We demonstrate this approach through a detailed case study focused on developing a user context profile ontology within the vehicle sales domain. Key contributions include significantly accelerated ontology construction processes, improved ontological consistency, effective bias mitigation, and enhanced transparency in the ontology engineering process. Our findings highlight the transformative potential of integrating LLMs into ontology development, notably improving scalability, integration capabilities, and overall efficiency in knowledge management systems.
Why we are recommending this paper?
Due to your Interest in Ontology for Products
Given the user's strong interest in knowledge graphs and ontology development, this paper's focus on leveraging LLMs for OKB creation is highly relevant. It offers a potential solution for scaling and improving knowledge management systems, a core area of interest.
Due to your Interest in Ontology for Products
Given the user's strong interest in knowledge graphs and ontology development, this paper's focus on leveraging LLMs for OKB creation is highly relevant. It offers a potential solution for scaling and improving knowledge management systems, a core area of interest.
Northwestern University
AI Insights
- The proof of Theorem 3.14 is based on the work of [ABD21, Theorem 3.7] and involves modifying their argument for the relative HHS case. [2]
Abstract
The Morse local-to-global property generalizes the local-to-global property for quasi-geodesics in a hyperbolic space. We show that graph products of infinite Morse local-to-global groups have the Morse local-to-global property. To achieve this, we generalize the maximization procedure of Abbott, Behrstock, and Durham for relatively hierarchically hyperbolic groups with clean containers. Under mild conditions satisfied by graph products, we show that stable embeddings into a relatively hierarchically hyperbolic space are exactly those which are quasi-isometrically embedded in the top level hyperbolic space by the orbit map. This shows that graph products of any infinite groups with no isolated vertices are Morse detectable.
The Morse local-to-global property generalizes the local-to-global property for quasi-geodesics in a hyperbolic space. We show that graph products of infinite Morse local-to-global groups have the Morse local-to-global property. To achieve this, we generalize the maximization procedure of Abbott, Behrstock, and Durham for relatively hierarchically hyperbolic groups with clean containers. Under mild conditions satisfied by graph products, we show that stable embeddings into a relatively hierarchically hyperbolic space are exactly those which are quasi-isometrically embedded in the top level hyperbolic space by the orbit map. This shows that graph products of any infinite groups with no isolated vertices are Morse detectable.
Why we are recommending this paper?
Due to your Interest in Graphs for Products
This paper explores graph products, a fundamental concept within knowledge graphs and their properties. The focus on Morse local-to-global properties aligns with the user's interest in graph-based representations of products and their relationships.
Due to your Interest in Graphs for Products
This paper explores graph products, a fundamental concept within knowledge graphs and their properties. The focus on Morse local-to-global properties aligns with the user's interest in graph-based representations of products and their relationships.
National University of Defense Technology
AI Insights
- The paper introduces a new task called Domain Knowledge Graph Fusion (DKGF), which aims to enhance the coverage and structural integrity of domain-specific knowledge graphs (DGKs) by leveraging high-quality general knowledge graphs (GKGs). [3]
- Two key challenges are identified in DKGF: high ambiguity of domain relevance and cross-domain knowledge granularity misalignment. [3]
- ExeFuse maps general facts to granularity-aware operators and verifies domain relevance through program executability, effectively integrating relevant and consistent knowledge from GKGs into DGKs. [3]
- Two benchmark datasets, DKGF(W-I) and DKGF(Y-I), are constructed to evaluate the performance of ExeFuse and other baseline models on the new task. [3]
- Domain Knowledge Graph Fusion (DKGF): a new task that aims to enhance the coverage and structural integrity of domain-specific knowledge graphs by leveraging high-quality general knowledge graphs. [3]
- General Knowledge Graphs (GKGs): high-quality knowledge graphs that contain general facts and relationships. [3]
- Domain-Specific Knowledge Graphs (DGKs): knowledge graphs that are specific to a particular domain or field. [3]
- Granularity-Aware Operators: operators that take into account the granularity of the knowledge being processed, such as entity-level or relation-level operations. [3]
- A Fact-as-Program paradigm called ExeFuse is developed to tackle the new task, reformulating DKGF as executable semantic reasoning over DGKs. [1]
Abstract
Domain-specific knowledge graphs (DKGs) often lack coverage compared to general knowledge graphs (GKGs). To address this, we introduce Domain-specific Knowledge Graph Fusion (DKGF), a novel task that enriches DKGs by integrating relevant facts from GKGs. DKGF faces two key challenges: high ambiguity in domain relevance and misalignment in knowledge granularity across graphs. We propose ExeFuse, a simple yet effective Fact-as-Program paradigm. It treats each GKG fact as a latent semantic program, maps abstract relations to granularity-aware operators, and verifies domain relevance via program executability on the target DKG. This unified probabilistic framework jointly resolves relevance and granularity issues. We construct two benchmarks, DKGF(W-I) and DKGF(Y-I), with 21 evaluation configurations. Extensive experiments validate the task's importance and our model's effectiveness, providing the first standardized testbed for DKGF.
Domain-specific knowledge graphs (DKGs) often lack coverage compared to general knowledge graphs (GKGs). To address this, we introduce Domain-specific Knowledge Graph Fusion (DKGF), a novel task that enriches DKGs by integrating relevant facts from GKGs. DKGF faces two key challenges: high ambiguity in domain relevance and misalignment in knowledge granularity across graphs. We propose ExeFuse, a simple yet effective Fact-as-Program paradigm. It treats each GKG fact as a latent semantic program, maps abstract relations to granularity-aware operators, and verifies domain relevance via program executability on the target DKG. This unified probabilistic framework jointly resolves relevance and granularity issues. We construct two benchmarks, DKGF(W-I) and DKGF(Y-I), with 21 evaluation configurations. Extensive experiments validate the task's importance and our model's effectiveness, providing the first standardized testbed for DKGF.
Why we are recommending this paper?
Due to your Interest in Knowledge Graphs
The paper's exploration of fusing domain-specific and general knowledge graphs directly addresses the user's interest in knowledge graph enrichment and expansion. This approach is crucial for building comprehensive and adaptable knowledge systems.
Due to your Interest in Knowledge Graphs
The paper's exploration of fusing domain-specific and general knowledge graphs directly addresses the user's interest in knowledge graph enrichment and expansion. This approach is crucial for building comprehensive and adaptable knowledge systems.
University of California, Los Angeles
AI Insights
- Fine-tuning: the process of adapting a pre-trained model to a specific task by adjusting its parameters. [3]
- The fine-tuned function vectors may not generalize well to new or unseen data. [3]
- The paper proposes a method for fine-tuning function vectors in large language models (LLMs) using an AdamW optimizer and L2 regularization. [2]
Abstract
Representing relations between concepts is a core prerequisite for intelligent systems to make sense of the world. Recent work using causal mediation analysis has shown that a small set of attention heads encodes task representation in in-context learning, captured in a compact representation known as the function vector. We show that fine-tuning function vectors with only a small set of examples (about 20 word pairs) yields better performance on relation-based word-completion tasks than using the original vectors derived from causal mediation analysis. These improvements hold for both small and large language models. Moreover, the fine-tuned function vectors yield improved decoding performance for relation words and show stronger alignment with human similarity judgments of semantic relations. Next, we introduce the composite function vector - a weighted combination of fine-tuned function vectors - to extract relational knowledge and support analogical reasoning. At inference time, inserting this composite vector into LLM activations markedly enhances performance on challenging analogy problems drawn from cognitive science and SAT benchmarks. Our results highlight the potential of activation patching as a controllable mechanism for encoding and manipulating relational knowledge, advancing both the interpretability and reasoning capabilities of large language models.
Representing relations between concepts is a core prerequisite for intelligent systems to make sense of the world. Recent work using causal mediation analysis has shown that a small set of attention heads encodes task representation in in-context learning, captured in a compact representation known as the function vector. We show that fine-tuning function vectors with only a small set of examples (about 20 word pairs) yields better performance on relation-based word-completion tasks than using the original vectors derived from causal mediation analysis. These improvements hold for both small and large language models. Moreover, the fine-tuned function vectors yield improved decoding performance for relation words and show stronger alignment with human similarity judgments of semantic relations. Next, we introduce the composite function vector - a weighted combination of fine-tuned function vectors - to extract relational knowledge and support analogical reasoning. At inference time, inserting this composite vector into LLM activations markedly enhances performance on challenging analogy problems drawn from cognitive science and SAT benchmarks. Our results highlight the potential of activation patching as a controllable mechanism for encoding and manipulating relational knowledge, advancing both the interpretability and reasoning capabilities of large language models.
Why we are recommending this paper?
Due to your Interest in Knowledge Management
This paper tackles the critical challenge of representing relationships within knowledge graphs, aligning with the user's interest in structured knowledge representation and reasoning. The use of function vectors offers a promising approach for capturing complex relational patterns.
Due to your Interest in Knowledge Management
This paper tackles the critical challenge of representing relationships within knowledge graphs, aligning with the user's interest in structured knowledge representation and reasoning. The use of function vectors offers a promising approach for capturing complex relational patterns.
University of
AI Insights
- The mutual-visibility number of a tree is equal to the number of its leaves. [2]
Abstract
The notion of mutual visibility in graphs arises from constraining shortest paths by forbidding internal vertices from belonging to a specified subset. Mutual-visibility sets, originally introduced as a tool for studying information flow and structural restrictions in complex networks, have since gained increasing attention due to their theoretical significance and diverse applications. In this paper, a complete characterization of mutual-visibility sets in trees is presented. It is shown that a subset $S$ is a mutual-visibility set of $T$ if and only if it coincides with the set of leaves of the Steiner subtree $T\langle S\rangle$. As a consequence, the mutual-visibility number of a tree is equal to the number of its leaves. For trees containing branch vertices, the notion of legs is introduced and an explicit formula for the number of maximal mutual-visibility sets is derived in terms of the corresponding leg lengths. It is proved that every tree is absolute-clear. It is further established that the mutual-visibility number is preserved under the line graph operation for trees with at least two edges, that is, $μ(L(T))=μ(T)$.
The notion of mutual visibility in graphs arises from constraining shortest paths by forbidding internal vertices from belonging to a specified subset. Mutual-visibility sets, originally introduced as a tool for studying information flow and structural restrictions in complex networks, have since gained increasing attention due to their theoretical significance and diverse applications. In this paper, a complete characterization of mutual-visibility sets in trees is presented. It is shown that a subset $S$ is a mutual-visibility set of $T$ if and only if it coincides with the set of leaves of the Steiner subtree $T\langle S\rangle$. As a consequence, the mutual-visibility number of a tree is equal to the number of its leaves. For trees containing branch vertices, the notion of legs is introduced and an explicit formula for the number of maximal mutual-visibility sets is derived in terms of the corresponding leg lengths. It is proved that every tree is absolute-clear. It is further established that the mutual-visibility number is preserved under the line graph operation for trees with at least two edges, that is, $μ(L(T))=μ(T)$.
Why we are recommending this paper?
Due to your Interest in Graphs for Products
Due to your Interest in Graphs for Products
University of California, San Diego
AI Insights
- The problem of computing the loss between two flows in a poset is reduced to calculating the supremum of the losses between certain pairs of elements. [2]
- The loss between two flows φ and ψ is defined as L(φ, ψ) = sup p≺qn Lp,q(φ), Lq △(φ, ψ), Lp,q(ψ), Lq ▽(φ, ψ)o . [1]
Abstract
The interleaving distance is arguably the most widely used metric in topological data analysis (TDA) due to its applicability to a wide array of inputs of interest, such as (multiparameter) persistence modules, Reeb graphs, merge trees, and zigzag modules. However, computation of the interleaving distance in the vast majority of this settings is known to be NP-hard, limiting its use in practical settings. Inspired by the work of Chambers et al. on the interleaving distance for mapper graphs, we solve a more general problem bounding the interleaving distance between generalized persistence modules on concrete categories via a loss function. This loss function measures how far an assignment, which can be thought of as an interleaving that might not commute, is from defining a true interleaving. We give settings for which the loss can be computed in polynomial time, including for certain assumptions on $k$-parameter persistence modules.
The interleaving distance is arguably the most widely used metric in topological data analysis (TDA) due to its applicability to a wide array of inputs of interest, such as (multiparameter) persistence modules, Reeb graphs, merge trees, and zigzag modules. However, computation of the interleaving distance in the vast majority of this settings is known to be NP-hard, limiting its use in practical settings. Inspired by the work of Chambers et al. on the interleaving distance for mapper graphs, we solve a more general problem bounding the interleaving distance between generalized persistence modules on concrete categories via a loss function. This loss function measures how far an assignment, which can be thought of as an interleaving that might not commute, is from defining a true interleaving. We give settings for which the loss can be computed in polynomial time, including for certain assumptions on $k$-parameter persistence modules.
Why we are recommending this paper?
Due to your Interest in Continual Generalized Category Discovery
Due to your Interest in Continual Generalized Category Discovery
Simon Fraser University
AI Insights
- The model is computationally expensive and not suitable for large datasets, but its performance using default settings is encouraging. [3]
- The authors have not compared their method with other recently-developed methods specifically for MLC prediction, which would be a useful extension. [3]
- The paper presents a Bayesian approach to multi-label classification using BART, which outperforms other methods in terms of accuracy and uncertainty quantification. [2]
Abstract
Multilabel Classification (MLC) deals with the simultaneous classification of multiple binary labels. The task is challenging because, not only may there be arbitrarily different and complex relationships between predictor variables and each label, but associations among labels may exist even after accounting for effects of predictor variables. In this paper, we present a Bayesian additive regression tree (BART) framework to model the problem. BART is a nonparametric and flexible model structure capable of uncovering complex relationships within the data. Our adaptation, MLCBART, assumes that labels arise from thresholding an underlying numeric scale, where a multivariate normal model allows explicit estimation of the correlation structure among labels. This enables the discovery of complicated relationships in various forms and improves MLC predictive performance. Our Bayesian framework not only enables uncertainty quantification for each predicted label, but our MCMC draws produce an estimated conditional probability distribution of label combinations for any predictor values. Simulation experiments demonstrate the effectiveness of the proposed model by comparing its performance with a set of models, including the oracle model with the correct functional form. Results show that our model predicts vectors of labels more accurately than other contenders and its performance is close to the oracle model. An example highlights how the method's ability to produce measures of uncertainty on predictions provides nuanced understanding of classification results.
Multilabel Classification (MLC) deals with the simultaneous classification of multiple binary labels. The task is challenging because, not only may there be arbitrarily different and complex relationships between predictor variables and each label, but associations among labels may exist even after accounting for effects of predictor variables. In this paper, we present a Bayesian additive regression tree (BART) framework to model the problem. BART is a nonparametric and flexible model structure capable of uncovering complex relationships within the data. Our adaptation, MLCBART, assumes that labels arise from thresholding an underlying numeric scale, where a multivariate normal model allows explicit estimation of the correlation structure among labels. This enables the discovery of complicated relationships in various forms and improves MLC predictive performance. Our Bayesian framework not only enables uncertainty quantification for each predicted label, but our MCMC draws produce an estimated conditional probability distribution of label combinations for any predictor values. Simulation experiments demonstrate the effectiveness of the proposed model by comparing its performance with a set of models, including the oracle model with the correct functional form. Results show that our model predicts vectors of labels more accurately than other contenders and its performance is close to the oracle model. An example highlights how the method's ability to produce measures of uncertainty on predictions provides nuanced understanding of classification results.
Why we are recommending this paper?
Due to your Interest in Product Categorization
Due to your Interest in Product Categorization
Seoul National University
AI Insights
- The 5W1H framework provides a universal performance lift across all LLMs regardless of the underlying prompt structure. [2]
- A QA scaffold effectively mitigates the complexity of raw text, but rigid entity-first filtering acts as a restrictive bottleneck. [1]
Abstract
Constructing Knowledge Graphs (KGs) from unstructured text provides a structured framework for knowledge representation and reasoning, yet current LLM-based approaches struggle with a fundamental trade-off: factual coverage often leads to relational fragmentation, while premature consolidation causes information loss. To address this, we propose SocraticKG, an automated KG construction method that introduces question-answer pairs as a structured intermediate representation to systematically unfold document-level semantics prior to triple extraction. By employing 5W1H-guided QA expansion, SocraticKG captures contextual dependencies and implicit relational links typically lost in direct KG extraction pipelines, providing explicit grounding in the source document that helps mitigate implicit reasoning errors. Evaluation on the MINE benchmark demonstrates that our approach effectively addresses the coverage-connectivity trade-off, achieving superior factual retention while maintaining high structural cohesion even as extracted knowledge volume substantially expands. These results highlight that QA-mediated semantic scaffolding plays a critical role in structuring semantics prior to KG extraction, enabling more coherent and reliable graph construction in subsequent stages.
Constructing Knowledge Graphs (KGs) from unstructured text provides a structured framework for knowledge representation and reasoning, yet current LLM-based approaches struggle with a fundamental trade-off: factual coverage often leads to relational fragmentation, while premature consolidation causes information loss. To address this, we propose SocraticKG, an automated KG construction method that introduces question-answer pairs as a structured intermediate representation to systematically unfold document-level semantics prior to triple extraction. By employing 5W1H-guided QA expansion, SocraticKG captures contextual dependencies and implicit relational links typically lost in direct KG extraction pipelines, providing explicit grounding in the source document that helps mitigate implicit reasoning errors. Evaluation on the MINE benchmark demonstrates that our approach effectively addresses the coverage-connectivity trade-off, achieving superior factual retention while maintaining high structural cohesion even as extracted knowledge volume substantially expands. These results highlight that QA-mediated semantic scaffolding plays a critical role in structuring semantics prior to KG extraction, enabling more coherent and reliable graph construction in subsequent stages.
Why we are recommending this paper?
Due to your Interest in Knowledge Graphs
Due to your Interest in Knowledge Graphs
Interests not found
We did not find any papers that match the below interests.
Try other terms also consider if the content exists in arxiv.org.
- MECE Mutually Exclusive, Collectively Exhaustive.
- Taxonomy of Products
You can edit or add more interests any time.
💬 Help Shape Our Pricing
We're exploring pricing options to make this project sustainable. Take 3 minutes to share what you'd be willing to pay (if anything). Your input guides our future investment.
Help us improve your experience!
This project is on its early stages your feedback can be pivotal on the future of the project.
Let us know what you think about this week's papers and suggestions!
Give Feedback